offline handwritten digit recognition using triangle ... · fingerprint recognition becomes...

International Journal of Computer Information Systems and Industrial Management Applications.

ISSN 2150-7988 Volume 10 (2018) pp. 087-097

© MIR Labs, www.mirlabs.net/ijcisim/index.html

Dynamic Publishers, Inc., USA

Received: 2 Jan, 2018, Accepted: 13 April, 2018, Publish: 23 April, 2018

Offline Handwritten Digit Recognition Using

Triangle Geometry Properties

Nur Atikah Arbain1*, Mohd Sanusi Azmi1, Azah Kamilah Muda1, Noor Azilah Muda1

and Amirul Ramzani Radzid1

1 Computational Intelligence and Technologies Lab,

Faculty of Information and Communication Technology,

Universiti Teknikal Malaysia Melaka, Hang Tuah Jaya,

76100 Durian Tunggal, Melaka, Malaysia

[email protected]

Abstract: Offline digit handwritten recognition is one of the

frequent studies that is being explored nowadays. Most of the

digit characters have their own handwriting nature. Recognizing

their patterns and types is a challenging task to do. Lately,

triangle geometry nature has been adapted to identify the pattern

and type of digit handwriting. However, a huge size of generated

triangle features and data has caused slow performances and

longer processing time. Therefore, in this paper, we proposed an

improvement on triangle features by combining the ratio and

gradient features respectively in order to overcome the problem.

There are four types of datasets used in the experiment which are

IFCHDB, HODA, MNIST and BANGLA. In this experiment, the

comparison was made based on the training time for each dataset

Besides, Support Vector Machine (SVM) and Multi-Layer

Perceptron (MLP) techniques are used to measure the accuracies

for each of datasets in this study.

Keywords: Digit Recognition, Handwriting, Triangle Feature,

Triangle Geometry.

I. Introduction

Offline handwriting recognition is handwriting that captured

optically via scanner and presented handwriting as an image.

In contrast, online handwriting recognition can be referred as a

method which implements an automatic processing using a

digitizer or any instrumented stylus that can capture any

information about the pen tip, for example, the position,

velocity or acceleration as a function of time [1].

Recently, research in offline digit recognition is often

explored because of interest in identifying the type, pattern and

origin of handwriting from various manuscripts. Most of the

manuscripts have their own handwriting nature and some of

them have used similar digit characters in different types of

manuscript [2]. For example, the Arabic characters have been

widely used in Jawi manuscripts [3]. Due to the researchers’

interest in offline digit recognition, it has given the opportunity

for them to explore and propose various techniques such as

Hidden Markov Model (HMM), Neural Network (NN) and

Triangular Block to recognize the handwriting.

However, not all techniques can be used to recognize the

digit handwriting. For example, the Chinese characters contain

a lot of strokes that differentiates the writers. Meanwhile, the

Arabic characters consist a lot of dots and critical marks in

sentences which contributed to huge challenges to the

researchers. Not only that, the Roman characters are also a

challenging handwriting to identify their physicality. A

suitable process is needed to extract the features because

slanted handwriting is hard to recognize. Thus, numerous

research and experiments have been conducted to produce

better accuracies in identifying the handwriting. In some cases,

the techniques will be combined and modified in order to

produce an appropriate approach to extract the features. This

is because the combination of techniques may produce a better

result of accuracy for digit handwriting. Nevertheless, not all

techniques are suitable to be combined due to the certain

difficulty of handwriting itself.

Over four decades ago, the studies in offline ROMAN digit

recognition for characters handwriting was explored [4]. In

meantime, no publicly was available for standard datasets that

can be used by the researchers. However, the development in

offline digit recognition was gone on a swift expansion in the

last decade. The Modified NIST dataset (MNIST) was known

as the largest dataset for ROMAN handwriting which was

established as a result of handwritten digit classification

competition that was held in summer of 1992 [5].

Besides that, HODA dataset is also known as a largest digit

dataset as well as MNIST dataset. However, HODA dataset is

a Farsi digit handwriting. It has contains binary images of

102,352 digits. The binary images were extracted from 12,000

registration forms where the forms were filled up by B. Sc. and

senior high school students [6]. The HODA and MNIST

dataset respectively are digit dataset that frequently used by

many researchers in their works. The studies of digit

recognition had grew speedily along with advance made on

prior methods and techniques [7]–[14].

Previously, a nature of triangle geometry has been

introduced to identify the digit handwriting [15]. The

Arbain, N.A et al. 88

normalization data has been used to overcome the issue as

stated in [15] where a big gap between triangle properties;

such as the angle value and gradient or ratio value. The gap has

affected the classification accuracies in digit recognition

handwriting. The MNIST dataset has been used by [15] as a

research dataset. The MNIST [5] dataset was known as a large

volume of digit images and most popular among other datasets

such as HODA dataset. According to [16], the number of

sample for each class in this database is known as the

non-uniform corresponding to real life distributions. The

standard datasets of MNIST and HODA have been widely

used in digit recognition handwriting and produced an

impressive results of classification accuracies [17].

The study of digit recognition handwriting yields an

extensive field dealing with various aspects of this difficult

task. According to [16], digit handwriting recognition is a

subclass of handwriting recognition problem and has become

very popular in recent years. A lot of processes are required

before the digit handwriting recognition can be identified such

as converting handwritings to grayscale images and binary

form, feature extraction and classification. At this point, the

selection of feature extraction is a crucial part to obtain high

digit handwriting recognition rate. The comprehensive review

for recent handwriting of digit recognition is discussed in [2],

[17]–[20].

This paper introduces the improvement on triangle

geometry features from the previous model as in [21]. The

experiments were conducted using two different machine

learning techniques which are the Support Vector Machine

(SVM) and the Multi-Layer Perceptron (MLP). The

multi-zoning method was used in feature extraction process as

similarly described in [22].

In this present paper, triangular block approach, problem

and data preparation are explained in Section II. Next, the

proposed method for triangle geometry features improvement

is discussed in Section III. The finding from an investigation is

discussed in Section IV and finally, the paper is concluded in

Section V.

II. Methodology

A. Triangular Block Approach

Triangular block approach has been widely used not only in

handwriting recognition but also in face recognition [23]–[25],

fingerprint recognition [26]–[28], vehicle detection [29]–[31]

and intrusion detection research [32], [33]. The triangle

geometry has become one of a prominent method to extract

features since the properties of triangle geometry can be

applied. The triangle sides are clustered into three types which

are equilateral, isosceles and scalene triangles, while for

triangle angle there are three types of angles which are right,

obtuse scalene and acute scalene triangles.

In face recognition, triangle points are acquired based on

body elements such as nasal tip, eyes, nose and mouth

[23]–[25]. The author of [23] has propose system for facial

recognition. The facial points were defined using elastic bunch

graph matching (EBGM) algorithm while

Kanade-Lucas-Tomaci (KLT) was used for tracking. The

geometric features were extracted from point, line and triangle

composed of tracking results of facial points [23]. The

architecture of the proposed facial expression recognition

system used by [23] is shown in Figure 1.

Figure 1. The architecture of proposed facial expression

recognition system [23]

While fingerprint recognition, triangle points are attained

based on minutiae [26]–[28]. The Delaunay triangulation is

one of popular method that has been widely used in

recognizing fingerprint. Based on [28], the triangulation can

be referred as the maximal planar subdivision whose vertex set

is P, where P denotes a finite set of points in a plane while

maximal planar subdivision was defined as a subdivision

where no edge connecting two vertices that can be added to the

subdivision without extinguishing the planarity. The Delaunay

triangulation method was used by [28] for fingerprint

verification. A modification for robust minutiae based

fingerprint verification was proposed by [28] where the

modification was for lessen the number of comparison

operations and the error rates within the matching process by

performing the full analysis of Delaunay triangulation. The

minutiae in [28] was represented by nodes of a coZnnected

graph composed of triangles. The example fingerprint image

using Delaunay triangulation is shown respectively in Figure 2

and Figure 3.

Figure 2. An example of fingerprint image 1 [28]

Offline Handwritten Digit Recognition Using Triangle Geometry Properties 89

Figure 3. An example of Delaunay Triangulation image 1 [28]

The triangle geometry method also has been used in

extracting features from digit images. The popular digit

datasets such as HODA [6] and IFCHDB [34] have been

extracted using various methods such as based on mixture of

RBF experts, Field Programmable Gate Array (FPGA),

decision templates method and local binary pattern [12],

[35]–[37]. In [35], the RBF experts was referred as a four RBF

neural network. The [35] has stated that the idea of the mixture

of experts method was based on the divide and conquer

principle where the complex problem was splitting into some

simple problems. Thus, the final result will be the mixture of

the small simple problem’s solutions. Besides, the loci

characterization method was applied for extracted features

through 45 and 135 degree directions. Based on [35], the loci

characterization feature vector for each image was determined

by placing a number to each background point in the image.

The example image of loci characterization feature is shown in

Figure 4.

Figure 4. A loci characterization feature for digit “seven”

[35]

The Field Programmable Gate Array (FPGA) is one of the

feature extraction method which has been applied by [12] for

offline Farsi handwritten digit recognition. The 11 features

(integer) were normalized into 40×40 pixel handwritten digit

images from HODA dataset. The block diagram of an FPGA

is shown in Figure 5 while division horizontal and vertical

section is shown in Figure 6.

Figure 5. The block diagram of system used in [12]

Figure 6. a) Divide into four horizontal section, b) Divide

into four vertical sections [12]

Nevertheless, the ways to determine three points of triangle

in face recognition and fingerprint recognition are different

compare to digit recognition. The triangle geometry used in

face and fingerprint cannot be applied on digit images due to

the constraints in digit images. Body elements are used to

determine triangle points in face recognition while minutiae

are used to determine triangle points in fingerprint. The ways

to determine triangle points used by face recognition and

fingerprint recognition becomes constraints to the digit images

because digit images do not have any body elements and

minutiae. Thus, the ways to determine three points of triangle

in digit recognition is using the proposed method from [21].

Based on [21], the centroid of image is used to

define point C of triangle based on the foreground colour

which is black. The centroid an image is given as:


and (1)

The point C of triangle divides image into two parts which is

left and right as shown in Figure 7.

Figure 7. Segregation process of binary image (1) [21]

Figure 8. Segregation process of binary image (2) [21]

The study in [22] has proposed nine features for offline digit

recognition. The multi-zoning method was proposed in

extracting triangle features into several parts. By using the

multi-zoning method, it is also used to generate number of

triangle features. The triangular block was incorporated into

zoning method. There are four types of zoning methods which

are Cartesian plane, Vertical plane, Horizontal plane and 45

degree-based zones.

The total zones produced by zoning method are 33 zones

which are also known as multi-zoning. Each of the zones will

generate nine triangle features that altogether produced 297

features from total triangle features of 33 zones. The

multi-zoning method has been discussed in [22]. The formula

to calculate the length of a, b and c are as shown in equation (2),

(3) and (4) while Figure 9 is an illustration of a triangle shape.

Table 1 shows the description of triangle shape in Figure 9

while Table 2 shows the triangle features description with

formula.

Figure 9. An illustration of triangle shape [15]

Table 1. Description of triangle shape

Corner Position Side Connected Angle

A Right b and c A

B Left a and b B

C Middle a and c C

(2)

(3)

(4)

Table 2. Description formula of triangle features [15]

No Feature Formula

1 c:a c:a = c/a

2 a:b a:b = a/b

3 b:c b:c = b/c

4 A

5 B

6 C

7 ΔBA

8 ΔBC

9 ΔCA


B. Problem in Processing Data

According to [10], accuracy and speed performance are the

essential parts that contributed to the whole performance of

digit recognition. In pattern classification and machine

learning groups, the problem of handwriting for digit

recognition is a good method to test the classification

performance [38]. The performance speed is influenced by the

large volume of data and the number of features.

A good technique for feature extraction plays important role

in data processing. Good feature extraction technique used

will contribute to a smooth and faster data processing even

though there is a numerous number of data.

In this paper, the discussed problem was based on research

in [21]. The total features produced are 297 features while a

total number of data for each dataset was more than 5000 data.

In data processing, a large volume of data will take longer time

for data extraction. This affected the performance during data

processing.

Thus, this study proposed the ideas of combining the ratio

and gradient features used in [21] in order to reduce the total

features and improve the performance during data processing.

C. Dataset Preparation

Four types of digit datasets are used in this study which are

Isolated Farsi/Arabic Character Database (IFCHDB) [34],

HODA [6], MNIST [5] and BANGLA [39].

The IFCHDB and HODA datasets are Arabic handwritings.

The MNIST dataset is one of the digit handwritings in Roman

while BANGLA dataset is one of the digit handwritings in

Indian language. Some datasets such as HODA and MNIST

can be downloaded freely from provided website. However,

IFCHDB and BANGLA datasets require the agreement form

to be filled in before requesting the samples data. After

completing the agreement form, samples data will be sent via

email. For HODA dataset, the samples data can be

downloaded from http://FarsiOCR.ir. For MNIST dataset, it

can be downloaded from http://yann.lecun.com/exdb/mnist/.

In this paper, each of the datasets is divided into two sets

which are the testing and the training data. Both testing and

training data contain 10 classes. These datasets have broad

characteristics and are employed as a benchmark for

improvement purposes.

The purpose of selecting these datasets was to confirm the

proposed algorithm with the characters which are not

originated from Arabic handwriting. That is why the MNIST

and BANGLA datasets are used in this experiment. Besides,

these datasets were chosen because, in this present study, the

result of accuracy and training time will be compared with the

previous investigation that has employed the same datasets.

Figure 10. Digit dataset used in the experiment

Table 3. Description of MNIST and BANGLA datasets

Dataset

Attribute

IFCHDB

[34]

HODA

[6]

Scale

Gray Binary

Training

12,292 56,790

Testing

5,268 20,000

Total 70,120 76,790

Table 4. Description of IFCHDB and HODA datasets

Dataset

Attribute

IFCHDB

[34]

HODA

[6]

Scale

Gray Binary

Training

12,292 56,790

Testing

5,268 20,000

Total 70,120 76,790

Table 5. An example of digit dataset images

IFCHDB HODA MNIST BANGLA

III. Proposed Method

A. An Overview of Proposed Method

In this paper, the proposed method is introduced to improve

the triangle features in [21] by combining the value of ratio

and gradient features based on zones. There are 33 zones used

in this paper. The 33 zones are using the zoning method to

extract the features.

Basically, there are three elements highlighted in this paper

which are the ratio, the gradient and the angle of the triangle.

The three main points of the triangle which are point A, B and

C are calculated based on the ratio, gradient and angle formula

in Table 2. Therefore, each of zones has generated nine


triangle features. After implementing the 33 zones, the total

number of features produced are 297 features.

In the proposed method, the ratio and gradient features are

the ratio for point A, B and C respectively while feature

number 7, 8 and 9 are the gradient for point A, B and C

respectively. In the proposed method, the values of feature

number 1, 2 and 3 are combined and become the main ratio.

For gradient, the values from feature 7, 8 and 9 are combined

and produced one feature known as the main gradient.

Therefore, the new total features for each zone are five

features. The angle of the triangle cannot be used because the

total value of triangle angle was 180 degree. Following are the

formula for calculating the main ratio and main gradient.

(5)

(6)

Based on Figure 12, the process is started by collecting the

digit dataset. The dataset contains raw images. After collecting

the digit dataset, these datasets are converted into a binary

form using Otsu threshold approach [40]. Once the digit image

conversion is completed, the zoning method that adopts the

nature of the triangle geometry is implemented in order to

extract the data. The zoning method is used in the feature

extraction stage. Next, the proposed method is implemented to

reduce the total features from 297 to 165 features. Lastly, the

final result with 165 features is produced.

In pre-processing stage, the triangle shape was performed

based on the coordinates for each of the zones. However, not

all triangle shape can be formed due to the collinear line

occurred. The collinear line was triggered because of the

coordinates for point A, B and C were too close to each other

or the gradient’s value produced is zero.

Thus, a detection on the coordinates of all triangle points

were catered using method in [17] which is also using triangle

geometry to solve the straight line problem in triangle shape

formation. The proposed method of [17] was focused on

detecting coordinates all points for each zones. Taking by the

example using Cartesian plane zone method, the partitions

were divided into five parts including main image. Using the

partition in Zone A (referred Figure 11), the partition of Zone

A was divided into four smaller parts. Then, the number of

pixel in each partitions of Zone A was calculated and

compared based on the rules stated in [17].

Figure 11. A process of proposed method to cater straight line problem [17]

Table 6. Description of features for proposed method

No of features Combination of non-related Features

1 (Ratio of point A) + (Ratio of point B) + (Ratio of point C) Main ratio of sides

2 (Gradient of point A) + (Gradient of point B) + (Gradient of point C) Main gradient of corner

3 Angle of point A Remain

4 Angle of point B Remain

5 Angle of point C Remain


Figure 12. The process of proposed method

B. Environmental Setup

The environment setup is a crucial part of processing the data.

There are two types of environment setup used. Table 7

presents the description of hardware used in this paper while

Table 8 presents the description of software used.

Table 7. Hardware description

Characteristic Item

Type of windows Windows 8

Processor Intel® Core™ i3-4160 CPU

@ 3.60GHz

Random Access Memory 12.0 GB

System Type 64-bit

Table 8. Software description

Software Description

Waikato Environment for

Knowledge Analysis

(WEKA)

Version 3.6.9

Eclipse Mars 1.0

Java Standard Edition Version 7 Update 76

IV. Result and Discussion

The experiment was conducted using Support Vector

Machine (SVM) and Multi-Layer Perceptron (MLP)

techniques. For SVM technique, the value of cost and gamma

were attained from grid search using the LIBSVM tool [41].

The result of cost and gamma for each dataset is shown in

Table 10. For MLP technique, the learning rate used was 0.3

which is obtained from the heuristic search. In this experiment,

all digit datasets aforementioned are used. First of all, the

comparison of classification accuracies is made between

several prior proposed method [36], [42] including [21].

However, among several proposed methods, the proposed

method of [21] is the only work that used the same triangle

geometry features as in this study. Table 9 shows the

comparison accuracy results from several previously proposed

methods.

Table 11 and Table 12 present the comparison of

classification accuracy results for each dataset between the

proposed method in [21] and our proposed method using SVM

and MLP techniques respectively. Table 13 and Table 14

present the results of comparing the training time taken for

each of the datasets between the proposed method in [21] and

our work.

Based on Table 11, the result of SVM technique accuracy

for our proposed method showed better outcome compared to

the proposed method in [21]. For IFCHDB dataset, the result

of SVM technique accuracy increases from 93.58% to 95.63%.

For HODA dataset, the result increases to 98.03% from

97.30% while BANGLA dataset result has increased from

90.28% to 93.29%. However, MNIST dataset showed value

reduction from 95.35% to 93.18%. The decreased result of

accuracy might be resulted from the nature of MNIST

handwriting itself.

For MLP technique, the accuracy in Table 12 for the

proposed method showed the unsatisfied result when

compared with the proposed method in [21]. This is due to the

complex calculation in MLP technique that may influence the

result accuracy. Besides, the nature of handwriting itself may

also influenced the result. However, the accuracy for MNIST

dataset has shown impressive result where the accuracy

increased from 88.66% to 96.51%. Therefore, the result of

accuracy as in Table 11 and Table 12 for our proposed method

showed better outcome compared to the method proposed

previously in [21].

Based on Table 13, the results of training time taken for

SVM technique were found to be faster than previous method

[21] after applying the proposed method. For IFCHDB dataset,

the result of training time taken was 88.89 seconds, 668.39

seconds for HODA dataset, 2397.38 seconds for MNIST

dataset and 194.49 seconds for BANGLA dataset. According

to Table 14, results of training time taken for MLP technique

showed improvement after using our proposed method. For

IFCHDB dataset, the result of training time taken was 1717.18

seconds, 5987.57 seconds for HODA dataset, 6198.89

seconds for MNIST dataset and 2295.55 seconds for

BANGLA dataset.

Overall, the results of our proposed method have shown

impressive accuracies with faster training time taken when

comparing with the proposed method in [21]. Therefore, this

study has proved that our proposed method has achieved the

target by improving the triangle features in [21] and adopting

the triangle geometry approach as in [21].

International Journal of Computer Information Systems and Industrial Management Applications.

ISSN 2150-7988 Volume 10 (2018) pp. 087-097

© MIR Labs, www.mirlabs.net/ijcisim/index.html

Dynamic Publishers, Inc., USA

Table 9. Comparison accuracy results from several prior proposed methods

Method IFCHDB HODA MNIST BANGLA

Characteristic Loci

and Principle

Component

Analysis [36]

MLP - 98.16 - -

Zoning, Outer

profiles, crossing

counts [42]

SVM - 98.90 - -

Triangle geometry

and zoning method

[21]

SVM

MLP

93.58

94.86

97.30

99.70

95.35

88.66

90.28

87.02

Table 10. Results of cost and gamma for each dataset

Dataset Cost (c) Gamma (𝜸)

IFCHDB 32.0 0.00048828125

HODA 8.0 0.001953125

BANGLA 32.0 0.001953125

MNIST 8.0 0.0078125

Table 11. Comparison of classification accuracy result for SVM technique (in %)


Triangle geometry and

zoning method [21]

93.58 97.30 95.35 90.28

Our proposed method 95.63 98.03 93.18 93.29

Table 12. Comparison of classification accuracy result for MLP technique (in %)



zoning method [21]

94.86 99.70 88.66 87.02


Table 13. Comparison of training time for SVM technique (in seconds)



zoning method [21]

112.87 1223.82 3703.21 231.77


Table 14. Comparison of training time for MLP technique (in seconds)



zoning method [21]

5188.68 18343.65 21751.96 6316.76



V. Conclusion

This paper presents the proposed method to improve triangle

features in [21] by combining the ratio and gradient (features

or characteristics). The ratio and gradient formula were

investigated and analyzed in order to produce a suitable

approach to improve the triangle features.

Focusing on digit recognition area, there were four datasets

used during the experiment. The accuracy result for each of the

datasets was measured by Support Vector Machine (SVM)

and Multi-Layer Perceptron (MLP) techniques. Besides, the

training time for each datasets was recorded and the result

were compared between the results in [21] and the present

findings. The training time result showed improvement in the

aspect of shorter time taken to process the data when

compared to the training time in [21]. Other than that, the

accuracies from the proposed method showed a better result.

However, the result of accuracy might be biased by the nature

of handwriting itself.

Feature extraction is an important factor in the performance

of character recognition. A powerful classifier such as SVM

and MLP may yield different accuracies based on different

patterns features. The high recognition accuracy can be

produced by selecting the suitable low-complexity classifier

and proper data extraction. The best final result is obtained not

only from the combination of the good classifier but also from

the features itself. The improvement in features helps to attain

high recognition accuracy and for sure the speed will be

increased with minimum training time taken. Further research

is needed to increase performance when constructing a triangle

shape.

Acknowledgment

The authors would like to express their appreciation to the

Universiti Teknikal Malaysia Melaka for providing a

scholarship of UTeM Zamalah Scheme. Besides, thank you to

the Universiti Teknikal Malaysia Melaka and Faculty of

Information Technology and Communication for providing

excellent research facilities.

References

[1] R. Plamondon and S. N. Srihari, “Online and off-line

handwriting recognition: a comprehensive survey,” IEEE

Trans. Pattern Anal. Mach. Intell., vol. 22, no. 1, pp.

63–84, 2000.

[2] A. N. Azmi, D. Nasien, and S. M. Shamsuddin, “A

review on handwritten character and numeral recognition

for Roman, Arabic, Chinese and Indian scripts,” Int. J.

Adv. Stud. Comput. Sci. Eng., vol. 2, no. 4, pp. 1–8,

2013.

[3] A. T. Gallop, “Beautifying Jawi: Between Calligraphy

and Paleography,” Second Int. Conf. Malay Civilis.

“Malay Image,” vol. I, pp. 194–233, 2005.

[4] A. L. Knoll, “Experiments with ‘Characteristic Loci’ for

Recognition of Handprinted Characters,” IEEE Trans.

Comput., vol. 18, no. April, pp. 366–372, 1969.

[5] A. Borji, M. Hamidi, and F. Mahmoudi, “Robust

handwritten character recognition with features inspired

by visual ventral stream,” Neural Process. Lett., vol. 28,

no. 2, pp. 97–111, 2008.

[6] H. Khosravi and E. Kabir, “Introducing a very large

dataset of handwritten Farsi digits and a study on their

varieties,” Pattern Recognit. Lett., vol. 28, no. 10, pp.

1133–1141, 2007.

[7] E. Kussul and T. Baidyk, “Improved method of

handwritten digit recognition tested on MNIST

database,” Image Vis. Comput., vol. 22, no. 12 SPEC.

ISS., pp. 971–981, 2004.

[8] K. Labusch, K. Labusch, E. Barth, E. Barth, T. Martinetz,

and T. Martinetz, “Simple Method for High-Performance

Digit Recognition Based on Sparse Coding,” Neural

Networks, IEEE Trans., vol. 19, no. 11, pp. 1985–1989,

2008.

[9] K. Labusch, K. Labusch, E. Barth, E. Barth, T. Martinetz,

T. Martinetz, R. M. O. Cruz, G. Cavalcanti, T. Ren, and J.

Malik, “Handwritten digit recognition using multiple

feature extraction techniques and classifier ensemble,”

Int. Conf. Syst. Signals Image Process., vol. 19, no. 11,

pp. 215–218, 2010.

[10] C. L. Liu, K. Nakashima, H. Sako, and H. Fujisawa,

“Handwritten digit recognition: Benchmarking of

state-of-the-art techniques,” Pattern Recognit., vol. 36,

no. 10, pp. 2271–2285, 2003.

[11] C.-L. Liu, K. Nakashima, H. Sako, and H. Fujisawa,

“Handwritten digit recognition: investigation of

normalization and feature extraction techniques,”

Pattern Recognit., vol. 37, no. 2, pp. 265–279, 2004.

[12] M. Moradi, M. A. Pourmina, and F. Razzazi,

“FPGA-Based farsi handwritten digit recognition

system,” Int. J. Simul. Syst. Sci. Technol., vol. 11, no. 2,

pp. 17–22, 2010.

[13] G. Mirsharif, M. Badami, B. Salehi, Z. Azimifar, and a

D. Collection, “Recognition of Farsi Handwritten Digits

Using a Small Feature Set,” Int. J. Comput. Electr. Eng.,

vol. 4, no. 4, pp. 588–591, 2012.

[14] V. Sundaresan and J. Lin, “Recognizing Handwritten

Digits and Characters,” 1998.

[15] M. S. Azmi, N. A. Arbain, A. K. Muda, Z. Abal Abas,

and Z. Muslim, “Data Normalization for Triangle

Features by Adapting Triangle Nature for better

Classification,” 2015 IEEE Jordan Conf. Appl. Electr.

Eng. Comput. Technol., pp. 1–4, 2015.

[16] M. Hamidi and A. Borji, “Invariance analysis of

modified C2 features: Case study-handwritten digit

recognition,” Mach. Vis. Appl., vol. 21, no. 6, pp.

969–979, 2010.

[17] N. A. Arbain, M. S. Azmi, S. S. S. Ahmad, I. E. A. Jalil,

M. Z. Masud, and M. A. Lateh, “Detection on Straight

Line Problem in Triangle Geometry Features for Digit

Recognition,” Int. J. Adv. Sci. Eng. Inf. Technol., vol. 6,

no. 6, pp. 1019–1025, 2016.

[18] S. Naz, S. Ahmed, R. Ahmad, and M. Razzak, “Arabic

Script based Digit Recognition Systems,” Int. Conf.


Recent Adv. Comput. Syst., no. 1, pp. 67–73, 2016.

[19] A. Lawgali, “Recognition of Handwritten Digits using

Histogram of Oriented Gradients,” Int. J. Adv. Res. Sci.

Eng. Technol., vol. 3, no. 7, pp. 2359–2363, 2016.

[20] P. K. Singh, R. Sarkar, and M. Nasipuri, “A Study of

Moment Based Features on Handwritten Digit

Recognition,” Appl. Comput. Intell. Soſt Comput., vol.

2016, pp. 1–17, 2016.

[21] M. S. Azmi, “Fitur Baharu Dari Kombinasi Geometri

Segitiga Dan Pengezonan Untuk Paleografi Jawi

Digital,” Doctoral dissertation, Universiti Kebangsaan

Malaysia, 2013.

[22] M. S. Azmi, K. Omar, M. F. Nasrudin, B. Idrus, and K.

Wan Mohd Ghazali, “Digit Recognition for Arabic/Jawi

and Roman using Features from Triangle Geometry,”

AIP Conf. Proc., vol. 1522, pp. 526–537, 2013.

[23] D. Ghimire, J. Lee, Z.-N. Li, and S. Jeong, “Recognition

of facial expressions based on salient geometric features

and support vector machines,” Multimed. Tools Appl., pp.

1–25, 2016.

[24] S. Soltanpour, B. Boufama, and Q. M. Jonathan Wu, “A

survey of local feature methods for 3D face recognition,”

Pattern Recognit., vol. 72, no. August, pp. 391–406,

2017.

[25] A. Shivdas, “Face Recognition Using Artificial Neural

Network,” Int. J. Res. Manag. Sci. Technol., vol. 2, no. 1,

pp. 61–65, 2014.

[26] E. M. Jadied, “Design and Implementation of Secure

Fingerprint Template Using Key Binding Scheme and

Random,” 2015 3rd Int. Conf. Inf. Commun. Technol.,

pp. 381–385, 2015.

[27] A. Jain and M. V. N. K. Prasad, “Clustering Based

Fingerprint Indexing Using Triangle Spiral,” 2015 11th

Int. Conf. Signal-Image Technol. Internet-Based Syst.,

pp. 81–88, 2015.

[28] M. Flores, G. Torres, G. García, and M. Licona,

“Fingerprint Verification Methods Using Delaunay

Triangulations,” vol. 14, no. 3, pp. 346–354, 2017.

[29] S. Parveen and F. Haque, “Traffic Sign Detection and

Recognition Using Colour Features towards Intelligent

Vehicle/Driver Assistance System,” J. Sci. Technol., vol.

2, no. 2, pp. 41–47, 2017.

[30] J. Kim, E. Yim, C. Jeon, C. Jung, and B. Han, “Cold

performance of various biodiesel fuel blends at low

temperature,” Int. J. …, vol. 13, no. 2, pp. 293–300,

2012.

[31] H. Zhu, K.-V. Yuen, L. Mihaylova, and H. Leung,

“Overview of Environment Perception for Intelligent

Vehicles,” IEEE Trans. Intell. Transp. Syst., pp. 1–18,

2017.

[32] P. S. Sarawagi, “Intrusion Detection Using Data

Mining,” Management, vol. 3, no. 4329022, pp. 93–101,

2017.

[33] P. Tang, R. Jiang, and M. Zhao, “Feature selection and

design olintrusion detection system based on k-means

and triangle area support vector machine,” Second Int.

Conf. Futur. Networks ICFN’10, pp. 144–148, 2010.

[34] S. Mozaffari, K. Faez, F. Faradji, M. Ziaratban, and S. M.

A Golzan, “ A Comprehensive Isolated Farsi/Arabic

Character Database for Handwritten OCR Research,”

Tenth Int. Work. Front. Handwrit. Recognit., pp.

385–389, 2006.

[35] R. Ebrahimpour, A. Esmkhani, and S. Faridi, “Farsi

handwritten digit recognition based on mixture of RBF

experts,” IEICE Electron. Express, vol. 7, no. 14, pp.

1014–1019, 2010.

[36] M. M. Javidi and F. Sharifizadeh, “A Modified Decision

Templates Method for Persian Handwritten Digit

Recognition,” J. Am. Sci., vol. 8, no. 1, p. 2012, 2012.

[37] M. Biglari, F. Mirzaei, and J. G. Neycharan,

“Persian/Arabic Handwritten Digit Recognition Using

Local Binary Pattern,” Int. J. Digit. Inf. Wirel. Commun.,

vol. 4, no. 4, pp. 486–492, 2014.

[38] Y. LeCun, L. Jackel, L. Bottou, a. Brunot, C. Cortes, J.

Denker, H. Drucker, I. Guyon, U. Müller, E. Säckinger,

P. Simard, and V. Vapnik, “Comparison of learning

algorithms for handwritten digit recognition,” Int. Conf.

Artif. Neural Networks, pp. 53–60, 1995.

[39] U. Bhattacharya and B. B. Chaudhuri, “Handwritten

numeral databases of Indian scripts and multistage

recognition of mixed numerals,” IEEE Trans. Pattern

Anal. Mach. Intell., vol. 31, no. 3, pp. 444–457, 2009.

[40] P. Smith, D. B. Reid, C. Environment, L. Palo, P. Alto,

and P. L. Smith, “A Threshold Selection Method from

Gray-Level Histograms,” IEEE Trans. Syst. Man.

Cybern., vol. 9, no. 1, pp. 62–66, 1979.

[41] C.-C. Chang and C.-J. Lin, “LIBSVM: A Library for

Support Vector Machines,” ACM Trans. Intell. Syst.

Technol., vol. 2, p. 27:1--27:27, 2011.

[42] M. Rajabi, N. Nematbakhsh, and S. Amirhassan

Monadjemi, “A New Decision Tree for Recognition of

Persian Handwritten Characters,” Int. J. Comput. Appl.,

vol. 44, no. 6, pp. 52–58, 2012.

Author Biographies

Nur Atikah Arbain was born in Melaka, Malaysia. She

received her Bachelor of Computer Science in Database

Management on 2015 and Master of Science in

Information and Communication Technology on 2016 at

Universiti Teknikal Malaysia Melaka. She is currently

pursuing her PhD which is also at the same university. Her

current research work is an offline subword handwriting

and contributes in feature extraction domain.

Mohd Sanusi Azmi received BSc., Msc and Ph.D from

Universiti Kebangsaan Malaysia (UKM) in 2000, 2003

and 2013. He joined Department of Software Enginering,

Universiti Teknikal Malaysia Melaka (UTeM) in 2003.

Now, he is currently a senior lecturer at UTeM. He is the

Malaysian pioneer researcher in identification and

verification of digital images of Al-Quran Mushaf. He is

also involved in Digital Jawi Paleography. He actively

contributes in the feature extraction domain. He has

proposed a novel technique based on geometry feature

used in Digit and Arabic based handwritten documents.

Azah Kamilah Muda is an Associate Professor at Faculty

of ICT, UTeM. She has appointed as Deputy Dean of Post

Graduate and Research since 2015. She received her PhD

in 2010 from Universiti Teknologi Malaysia, specializing

in image processing. Her research interest includes

fundamental studies on data analytics using soft

computing techniques, pattern analysis and recognition,

image processing, machine learning, computational

intelligence and hybrid systems. Her current research


work is on pattern analysis of molecular computing for

drug analysis, data analytic for various application and

root cause analysis in manufacturing process.

Noor Azilah Muda is a Senior Lecturer and a researcher

with 15 years of teaching and research experience in the

Software Engineering, Database Management and Soft

Computing area. Her research interest includes

fundamental studies on recognizing patterns of music

features and images besides involving in data related

research area such as data knowledge and data analysis.

Some of her works are in the area of Artificial Immune

System (AIS) mechanism to recognize cells that she later

investigates the mechanism to recognize different patterns

of music features for music genre classification.

Amirul Ramzani Radzid received the BSc. in Computer

Science (Software Development) from University

Teknikal Malaysia Melaka (UTeM) in 2016. Currently he

is pursuing Msc in Information and Communication

Technology at Universiti Teknikal Malaysia Melaka

(UTeM). His research interests include the area of image

segmentation.

offline handwritten digit recognition using triangle ... · fingerprint recognition becomes...

Documents